[PATCH v3] ARM: document arm_setup_stack() pitfalls

Ahmad Fatoum a.fatoum at pengutronix.de
Fri Sep 17 05:11:52 PDT 2021

Many arm32 board entry points use arm_setup_stack() to set up
the stack from C code. This necessitates using __naked, which
probably has been our most frequent cause of misscompiled C code.

GCC is quite clear that:

  Only basic asm statements can safely be included in naked functions
  While using extended asm or a mixture of basic asm and C code may
  appear to work, they cannot be depended upon to work reliably and
  are not supported.

But some boards use it anyway, because it's nice to avoid writing
assembly. Reading generated assembly to spot compiler miscompilation
isn't that nice though, so add some documentation, comments
and compiler diagnostics to hopefully reduce future porting effort.

Signed-off-by: Ahmad Fatoum <a.fatoum at pengutronix.de>
v2 -> v3:
  - fix typos (Uwe)
  - use better matching CONFIG_CPU_64 instead of CONFIG_CPU_V8
  - clarify 'bad' stack manipulation
v1 -> v2:
  - fix commit message title

Cc: Jules Maselbas <jmaselbas at kalray.eu>
 Documentation/devel/porting.rst | 21 +++++++++++++++++++++
 arch/arm/include/asm/common.h   | 17 +++++++++++++++++
 2 files changed, 38 insertions(+)

diff --git a/Documentation/devel/porting.rst b/Documentation/devel/porting.rst
index 97b787327c9a..1abaabc03ae1 100644
--- a/Documentation/devel/porting.rst
+++ b/Documentation/devel/porting.rst
@@ -169,6 +169,27 @@ Looking at other boards you might see some different patterns:
    needs to be done at start. If a board similar to yours does this, you probably
    want to do likewise.
+ - ``__naked``: All functions called before stack is correctly initialized must be
+   marked with this attribute. Otherwise, function prologue and epilogue may access
+   the uninitialized stack. If the compiler for the target architecture doesn't
+   support the attribute, stack must be set up in non-inline assembly:
+   Either a barebox assembly entry point or in earlier firmware.
+   The compiler may still spill excess local C variables used in a naked function
+   to the stack before it was initialized.
+   A naked function should thus preferably only contain inline assembly, set up a
+   stack and jump directly after to a ``noinline`` non naked function where the
+   stack is then normally usable.
+ - ``noinline``: Compiler code inlining is oblivious to stack manipulation in
+   inline assembly. If you want to ensure a new function has its own stack frame
+   (e.g. after setting up the stack in a ``__naked`` function), you must jump to
+   a ``__noreturn noinline`` function.
+ - ``arm_setup_stack``: For 32-bit ARM, ``arm_setup_stack`` initializes the stack
+   top when called from a naked C function, which allows to write the entry point
+   directly in C. The stack pointer will be decremented before pushing values.
+   Avoid interleaving with C-code. See ``__naked`` above for more details.
  - ``__dtb_z_my_board_start[];``: Because the PBL normally doesn't parse anything out
    of the device tree blob, boards can benefit from keeping the device tree blob
    compressed and only unpack it in barebox proper. Such LZO-compressed device trees
diff --git a/arch/arm/include/asm/common.h b/arch/arm/include/asm/common.h
index d03ee6273fe5..1e2729d521ae 100644
--- a/arch/arm/include/asm/common.h
+++ b/arch/arm/include/asm/common.h
@@ -1,6 +1,8 @@
 #ifndef __ASM_ARM_COMMON_H
 #define __ASM_ARM_COMMON_H
+#include <linux/compiler.h>
 static inline unsigned long get_pc(void)
 	unsigned long pc;
@@ -46,8 +48,23 @@ static inline unsigned long get_sp(void)
 	return sp;
+extern void __compiletime_error(
+	"arm_setup_stack() called outside of naked function. On ARM64, "
+	"stack should be setup in non-inline assembly before branching to C entry point."
+) __unsafe_setup_stack(void);
+ * Sets up new stack growing down from top within a naked C function.
+ * The first stack word will be top - sizeof(word).
+ *
+ * Avoid interleaving with C code as much as possible and jump
+ * ASAP to a noinline function.
+ */
 static inline void arm_setup_stack(unsigned long top)
+		__unsafe_setup_stack();
 	__asm__ __volatile__("mov sp, %0"
 			     : "r"(top));

More information about the barebox mailing list